A comparison of prostate cancer bone metastases on 18F-Sodium Fluoride and Prostate Specific Membrane Antigen (18F-PSMA) PET/CT: Discordant uptake in the same lesion

Purpose Prostate-Specific Membrane Antigen (PSMA) PET/CT has been introduced as a sensitive method for characterizing metastatic prostate cancer. The purpose of this study is to compare the spatial concordance of 18F-NaF PET/CT and 18F-PSMA-targeted PET/CT within prostate cancer bone metastases. Methods Prostate cancer patients with known bone metastases underwent PSMA-targeted PET/CT (18F-DCFBC or 18F-DCFPyL) and 18F-NaF PET/CT. In pelvic and spinal lesions detected by both radiotracers, regions-of-interest (ROIs) derived by various thresholds of uptake intensity were compared for spatial colocalization. Overlap volume was correlated with uptake characteristics and disease status. Results The study included 149 lesions in 19 patients. Qualitatively, lesions exhibited a heterogeneous range of spatial concordance between PSMA and NaF uptake from completely matched to completely discordant. Quantitatively, overlap volume decreased as a function of tracer intensity. and disease status, where lesions from patients with castration-sensitive disease showed higher spatial concordance while lesions from patients with castration-resistant disease demonstrated more frequent spatial discordance. Conclusion As metastatic prostate cancer progresses from castration-sensitive to castration-resistant, greater discordance is observed between NaF PET and PSMA PET uptake. This may indicate a possible phenotypic shift to tumor growth that is more independent of bone remodeling via osteoblastic formation.


INTRODUCTION
Most men who develop metastatic prostate cancer in the bone inevitably succumb to the disease [1]. It is thought that once seeding of prostate cancer cells within the bone has occurred and formation of metastasis begins, continued growth and proliferation of cancer cells is dependent on a complex interaction between the cancer cells and the bone microenvironment, leading to a 'vicious cycle' of osteoblast recruitment and osteoclastic response within the local area of metastatic invasion [2]. This eventually leads to the osteosclerotic lesion associated with prostate cancer metastases. Radionuclide bone scans have long been regarded as a reliable modality for imaging bone metastases due to their sensitivity in detecting regions of high bone turnover. Several types of bone scans are currently available including 99m Tc methylene diphosphonate (MDP) planar bone scans, which are the traditional method, and relatively newer techniques such as 99m Tc-MDP SPECT/CT and 18 F-NaF PET/CT which show superior detection sensitivity [3,4]. These agents are all taken up in areas of active bone remodeling, specifically new bone formation indicating increased osteoblast activity by osteocytes that are actively laying down bone.
Recently, prostate cancer imaging agents targeting Prostate Specific Membrane Antigen (PSMA), which is highly expressed in prostate cancer cells, have become available. PSMA contains a carboxypeptidase that is thought to cleave glutamate from vitamin B9, possibly activating the phosphoinositide 3-kinase pathway and promoting prostate cancer growth independent of the androgen receptor (AR) pathway [5]. Several PSMAtargeted radiotracers are in clinical trials evaluation, all of which bind to the enzymatic region of PSMA with high affinity and target viable tumor providing a unique biomarker for metastatic disease [6][7][8].
Previous reports have indicated that there are discrepancies in detection of bone metastases between bone imaging agents and PSMA agents in metastatic prostate cancers [9][10][11][12][13][14]. These discrepancies are reported at the lesion level, i.e. a lesion is seen either on one but not the other. This may depend on disease status (PSMA expressing or not) and the effects of therapy such as suppression of the AR pathway with androgen deprivation therapy (ADT) [12]. However, so far little attention has been paid to spatial colocalization of 18 F-NaF PET/CT and PSMA PET/CT uptake within the same bone lesion. While the microenvironment within prostate cancer bone metastases contains both osteocytes and cancer cells, the presence and degree to which they co-localize has yet to be functionally evaluated across individual lesions. Given the different targeting mechanisms it is feasible uptake distributions may be different within the same lesion, and findings of spatial discordance have potential therapeutic implications for patients with boney prostate metastases. In this study, uptake by both PET/CT agents within individual lesions were carefully mapped by CT coregistration with the purpose to characterize the degree of spatial concordance between 18 F-NaF PET and 18 F-PSMA PET uptake.

RESULTS
Fifty-two patients with metastatic prostate cancer were enrolled across both imaging studies. Nineteen patients met study inclusion criteria with bone metastases in the pelvis or spine co-detected by NaF PET/CT and PSMA PET/CT. Patients were excluded based on soft tissue only disease (N=9), patients with NaF-only bone lesions (N=14), patients with negative scans (N=7), or patients with co-detected lesions not meeting inclusion criteria (N=6). A summary of patient demographics and lesion characteristics are listed in Table 1, with patientspecific characteristics listed in Supplementary Table 1. 224 prostate cancer bone metastases were detected by both NaF and PSMA imaging. 149 of these lesions met inclusion criteria for analysis. Qualitatively, metastases visually exhibited various degrees of matching between the two scan types, as summarized in Figures 1-3. While some lesions demonstrated nearly complete concordance ( Figure 1A, 1B) of uptake, others showed regionallysimilar distribution manifested as moderate ( Figure 2A) or high ( Figure 2B) concordance. This further devolved into patterns where partial ( Figure 3B) to substantial ( Figure  3C) spatial uptake discordance was observed.
Robust CT registration between NaF PET/CT and PSMA PET/CT imaging studies allowed robust registration as was demonstrated by an ICC of 0.926. In the mixed effects model of CT characteristics, Hounsfield Unit (HU) values were shown to vary regionally within voxels categorized by which types of ROIs they were contained within ( Table 2). Voxels contained only within PSMA ROIs demonstrated the lowest HU, indicating less sclerosis, while voxels contained only within CT ROIs demonstrated the highest HU (Supplementary Figure 1). Only one lesion in this patient population was characterized by true osteolytic appearance, demonstrating poor spatial colocalization (Supplemental Figure 2). Similar to the qualitative findings, concordance of NaF and PSMA ROI volumes varied substantially across all lesions, with median overlap volume 0.77 (range 0-1). PSMA ROIs showed higher concordance with NaF ROIs compared to CT ROIs (p=0.047), while NaF volumes showed similar spatial overlap with CT (median 0.75, range 0-1) as observed with NaF and PSMA volumes (Supplementary Figure 3). Lesions within the same patient showed varying degrees of spatial discordance and patient-level heterogeneity was observed across metastatic burden (Supplementary Figure 4).
ROI volumes of each bone lesion were further segmented at multiple uptake intensity levels, as described in Figure 4. In this analysis, overlap between NaF and PSMA volumes decreased incrementally at image intensity thresholds of 60%-SUV max (p=0.002), 70%-SUV max (p=0.0005), and 80%-SUV max (p=0.0008). This finding varied according to whether patients were considered to have castration-sensitive (N=6) vs. castration-resistant (N=13) disease, as shown in Figure  5. Metastatic bone lesions in patients with castratesensitive disease maintained moderate spatial concordance across all intensity threshold levels (median 62.5% in 80%-SUV max ROIs) while the degree of overlap in lesions within patients with castration-resistant disease decreased significantly at each intensity level (median 21.5% in 80%-SUV max ROIs). Overlap volume in the castrationsensitive group was modestly higher than the castrateresistant group at 60%-SUV max (p=0.15), becoming significant at 70%-SUV max (p=0.02) and 80%-SUV max (p=0.004). Distance metrics between areas of highest uptake in NaF and PSMA additionally show higher separation (more discordance) in CRPC patients (Table 3). Correlation between overlap volume, PET uptake from both radiotracers, and CT characteristics are summarized in Supplementary Table 2. Overlap volume showed weak-to-no correlation to NaF ROI volume (ρ=0.20) and PSMA ROI volume (ρ=0. 29), remaining so at all levels of intensity segmentation. Weak-to-moderate inverse correlation was observed for overlap volumes at all image segmentation levels vs. SUV max of NaF, ranging novo metastatic disease (serum PSA 16 ng/ml) demonstrating 100% volumetric overlap between 18 F-NaF uptake (top row, blue contour) and 18 F-PSMA (DCFPyL) (bottom row, red contour) uptake, with CT contour shown in green. (B) Lesion in sacrum of castrate-sensitive patient (serum PSA 3.05 ng/ml) demonstrating 88.7% volumetric overlap between 18 F-NaF uptake (top row, blue contour) and 18 F-PSMA (DCFPyL) (bottom row, red contour) uptake, with CT contour shown in green. In both (A) and (B), top row demonstrates PET uptake, PET/CT overlay, and low-dose CT from 18 F-NaF scan and bottom row demonstrates PET uptake, PET/CT overlay, and low-dose CT from 18 F-PSMA (DCFPyL) scan with registered contour overlay. www.oncotarget.com from -0.09 to 0.23, and PSMA, ranging from -0.03 to -0.29 (Supplementary Table 2). A modest, association between HU mean and overlap volume was observed at all PSMA segmentation levels (ranging from ρ=0.31 to 0.36).
No correlation was demonstrated for SUV max and HU mean between PET imaging agents and any segmentation level (Supplementary Table 3).

Figure 2: Examples of moderate-to-high volume concordance, with differential localization of highest uptake. (A)
Lesion in anterior iliac crest of castrate-resistant patient (serum PSA 93.3 ng/ml) demonstrating moderate 74.9% volumetric overlap between 18 F-NaF uptake (top row, blue contour) and 18 F-PSMA (DCFPyL) (bottom row, red contour) uptake, but mismatching patterns of high uptake within contours and areas of scelrosis on CT (green contour). (B) Lesion in T12 spine of castrate-resistant patient (serum PSA 388.1 ng/ml) demonstrating high 95.5% volumetric overlap between 18 F-NaF uptake (top row, blue contour) and 18 F-PSMA (DCFBC) (bottom row, red contour) uptake within visible CT lesion (green contour), but differential areas of highest uptake. In both (A) and (B), top row demonstrates PET uptake, PET/CT overlay, and low-dose CT from 18 F-NaF scan and bottom row demonstrates PET uptake, PET/CT overlay, and low-dose CT from PSMA scans with registered contour overlay. www.oncotarget.com

DISCUSSION
This study demonstrates that many prostate bone metastases are moderately to highly discordant in regard to regions of uptake of 18 F-NaF PET and 18 F-PSMA. It is well established that prostate cancer bone metastases induce an osteoblastic bone reaction, visible on bone scans and CT imaging [15]. However, this does not imply perfect concordance between areas of high bone turnover and areas of highest cancer activity. Indeed, it is a common experience that biopsies of sclerotic or bone scan-positive lesions for pathological confirmation of metastatic disease are often negative [16]. While the presence of enhanced bone activity and formation is known to exist in metastatic sites, the spatial concordance with cancer cells has not been functionally evaluated. In this analysis, we report a highly variable spatial concordance of NaF and PSMA PET uptake within the same bone metastases, in which the degree of discordance increases as the disease progresses to castration-resistance.
The disturbance of normal bone homeostasis and active remodeling in sites of skeletal metastatic disease occurs early in the prostate cancer metastatic process and persists throughout the disease, resulting in predominately osteoblastic bone lesions [2,17]. This knowledge has led to use of the standard MDP bone scan and 18 F-NaF PET/ Region of overlap was assigned as one of 7 possible categorical memberships for each voxel, using voxels contained within all ROIs (PSMA, NaF, CT) as reference for comparison. demonstrates moderate spatial concordance of ROI volumes in right sacral bone lesion of castration-resistant patient (serum PSA 4379 ng/ ml) imaged with 18 F-PSMA (DCFBC) and 18 F-NaF. For each radiotracer, additional ROIs are then derived from voxels within gradientbased volumes that fall within 60% of SUV max , 70% of SUV max , and 80% of SUV max . Thus, each ROI is incrementally smaller in volume and higher in uptake relative to SUV max of each tracer within each bone lesion. The example provided demonstrates spatially discordant regions of highest uptake, with decreasing areas of ROI overlap at increasing levels of relative tracer intensity. www.oncotarget.com CT for patient assessment. These imaging agents target areas of increased bone turnover and new bone formation [18]. Spatially, this interaction of prostate cancer cells with the bone microenvironment coincides, such that the region of active bone remodeling usually reflects the region of active cancer [17]. This theory is supported by several findings that tumor burden in a metastatic lesion is regulated by tumor cell interactions with cells within the bone microenvironment, which includes osteoblasts, osteoclasts, and bone stromal cells [19][20][21][22]. We find high spatial concordance of osteoblastic activity and cancerspecific PSMA activity occurs more frequently in early (castrate-sensitive) disease. Lesions in patients with advanced castrationresistant disease [23], show more discordance between regions of osteoblastic activity and PSMA positive tissue. These findings could support the theory that the heterogeneous nature of metastatic prostate cancer, while driven by osteoblastic response initially, eventually evolves to become independent of an osteoblastic response [20], leading to dynamic pathobiologic diversity in endstage disease [24]. High PSMA PET uptake in areas suspicious for metastasis is considered highly reliable for prostate cancer and discordance with NaF can be interpreted to suggest there is an invasive component not driven by osteoblastic reaction. This parallels bone findings in metastatic breast cancer, where predominately osteolytic mechanisms are suspected to underlay observed discordance between NaF PET/CT and FDG PET/CT [25,26].
These findings have potential clinical relevance with regard to recently developed therapies for metastatic prostate cancer. 223 Ra has been approved for treatment of metastatic bone disease [27]. Preliminary studies show spatial correlation between 18 F-NaF PET/CT uptake and 223 Ra deposition [28]. Since 223 Ra is an alpha emitter its region of therapeutic efficacy is narrowly defined (within only a few cell diameters), Use in patients with a high degree of spatial discordance between 18 F-NaF and 18 F-PSMA could potentially result in undertreatment of disease within these lesions.
Different pairs of PSMA PET and bone scan agents have shown differing detection performance of metastatic bone disease in the past [11][12][13]. Similar detection differences with other PET agents used in metastatic prostate cancer have been reported [29][30][31]. These findings are likely the result of complex interactions between treatment timing/effect and disease state, underscoring the need for reliable imaging biomarkers throughout the course of disease. In the absence of histological validation for all lesions, inclusion of lesions detected by both tracers allows for phenotypic comparisons within lesions most likely to contain metastatic prostate cancer. Several studies indicate that CT appearance and density correlates to differences in uptake between various radiotracers [31,32], demonstrating NaF has a higher tumor-to-background ratio in osteosclerotic lesions [13]. Our voxel-based analysis are in general agreement, demonstrating regions defined only by PSMA uptake associate with regions of lower HU.
This study has several limitations. Accurate image segmentation for bone lesions remains an ongoing issue in PET/CT. To account for this fact, multiple image intensity based segmentations were performed in this study to investigate colocalization of tracer uptake without dependency on tumor volume or extent. No partial volume corrections were made, though both cohorts were reconstructed using time-of-flight algorithms, which in combination with point-spread function modeling, may limit the influence of partial volume effects [33]. It is unclear the presence or density of tumor cells and osteoid components within these phenotypic regions. The resolution of PET limits the detection capabilities to the extent that could be achieved in histopathological analysis; however, tissue-based sampling in all metastatic sites is infeasible. Furthermore, the influence of treatment status at the time of imaging on the colocalization of tracers could not be evaluated in this small cohort. A disproportionate number of castration-sensitive patients were in the DCFPyL cohort (6/7); however, similar rates of decreasing overlap volume were observed between imaging cohorts (Supplementary Figure 5). Longitudinal assessment at multiple timepoints by these tracers would be helpful to understand the dynamics of disease activity and tracer uptake over time.
In conclusion, 18 F-NaF and 18 F-PSMA PET scans for metastatic prostate cancer can show considerable discordance in regions of increased uptake within a bone metastasis. The relationship between PSMA activity in prostate metastases and bone turnover appears to become weaker in more advanced stages of disease, with metastases occupying portions of bone with no apparent bone turnover or osteoblastic proliferation. In addition to possibly providing insights into the evolution of prostate cancer metastatic spread, these findings have potential implications for radionuclide therapies such as 223 Ra that depend on localization in areas of increased bone turnover.

Patient population
This study population includes patients with known metastatic prostate cancer undergoing PSMAtargeted PET/CT, using either 18 F-DCFBC PET/ CT imaging (NCT02190279) or 18 F-DCFPyL PET/ CT imaging (NCT03173924), and NaF PET/CT for metastatic disease assessment at a single institution. These studies were approved by the Institutional Review Board and were Health Insurance Portability and Accountability Act compliant. All patients enrolled after written informed consent was obtained. Eligibility required histopathologically confirmed prostate cancer and identifiable metastatic disease on conventional imaging (CT, magnetic resonance imaging or bone scan). All patients received 18 F-DCFBC PET/CT or 18 F-DCFPyL PET/CT and 18 F-NaF PET/CT scans. Clinical demographics, including castration sensitivity/resistance status and prior treatment history were established based on clinical review of patient medical records.

Lesion-based inclusion criteria and image analysis
Standardized Uptake Values (SUVs) were calculated as the ratio of measured activity to injected dose per body weight (kilogram). Image review and analysis was performed using commercial software (MIM v.6.6.10, Cleveland, OH, USA). Lesions determined to be highly suspicious for metastatic disease by consensus of 3 nuclear medicine physicians were considered for analysis. Of highly suspicious bone lesions detected both by NaF PET/CT and PSMAtargeting PET/CT, only those within the pelvis and spine were included in spatial analysis to reduce artifacts introduced from breathing motion (ribs) or patientpositioning motion (extremities, skull). Regions-ofinterest (ROI) for each tracer were obtained by semiautomated gradient-based method (PETEdge, MIM). ROIs were further segmented into areas of highest tracer uptake by 60%, 70% and 80% thresholds of lesionspecific SUV max, as described in Figure 4. Regions corresponding to PET-detected lesions were segmented by CT appearance, i.e. encompassing radiographic visibility, by nuclear medicine specialists.
A two-step image registration procedure was completed using low dose CTs. First global skeletal alignment was achieved using rigid alignment. Next, lesion-based alignment was further achieved using local box-based optimization fit to bony regions containing each lesion (Box-based Assisted Alignment, MIM). After registration, spatial concordance of NaF and PSMA ROIs was evaluated by assessing the degree of overlap volume in the area of increased radiotracer uptake, defined as the ratio of overlapping volume to minimum lesion volume. This calculation was repeated for NaF and PSMA overlap with CT. Each voxel was assigned to one of seven possible concordance categories: PSMA exclusive, NaF exclusive, CT exclusive, PSMA and NaF only, PSMA and CT only, NaF and CT only, or all matching (included in all PSMA, NAF, and CT). Distance between regions of highest tracer uptake was measured by, absolute distance from PSMA SUV max and NaF SUV max . To avoid potential bias from partial-volume errors, the distance between center-of-mass and average pairwise distance from 80%-SUV max PSMA ROIs and 80%-SUV max NaF ROIs were also calculated.

Statistical analysis
Voxel-based accuracy of lesion-specific registration was evaluated using the Intraclass Correlation Coefficient (ICC), estimated from a mixed effect model of Hounsfield Units (HU) measurements with nested random effects for patient, lesions, skeletal region and PSMA tracers (DCFBC or DCFPyL) and fixed effect voxel-based concordance category. The significance of fixed effects was determined by the likelihood ratio test. Differences in lesion-based overlap volume as a function of ROI segmentation method was evaluated using paired Wilcoxon signed-rank test using Rosner-Glynne-Lee method to account for intrapatient correlation [36]. Association of overlap volume and distance metrics with disease status (castrationsensitive vs. castration-resistant) was evaluated by permutation test with 2000 permutations at the patient level. Correlation between SUV metrics, PET tumor uptake volume, and CT characteristics were completed using the Spearman correlation coefficient. Standard errors and 95% confidence intervals (CI) were estimated from 2000 bootstrap samples by random sampling on the patient-level. All p-values correspond to two-sided tests, with a p-value <0.05 considered to represent a significant difference between results.